Textual Analysis of Banks' Pillar 3 Documents

Abstract

Using a sample of 188 European listed banks covering 2004 to 2016, we conduct textual analysis on banks' Pillar 3 reports and annual reports to showcase how banks formulate their regulatory reports. We first develop dictionaries relying on machine learning tools and its subfield of textual analysis. In addition, we construct measures of text complexity (FOG), text detail (Named Entity Recognition), and boilerplate. We validate our dictionaries by extracting numerous textual indices consisting of (1) sentiment scores, (2) measures of tone, (3) similarity of reports over time, and (4) similarity across banks. As an additional validation test, we examine how markets react to the textual difference in the Pillar 3 disclosure of riskiness and uncertainty, proxied by sense, tone, and preciseness.

Keywords: Sentiment analysis, NLP, Text analysis, Accounting, Financial institution, Bank, Pillar 3

JEL Classification: D82, D83, G14, G18, G30, M40, M41